Category · Voice & speech
Speech-to-text APIs for AI agents
APIs that turn audio into text, streaming or in batch. Compared on accuracy across accents, noise, names and overlapping speakers, streaming latency, languages, diarisation and cost per audio minute.
Best speech-to-text APIs for AI agents, the shortlist387 head-to-head comparisons
Capability keys speech.stt · speech.streaming · speech.batch · speech.diarisation · speech.languages · All tools
letme.dev/speech.stt picks the top-graded tool in this list and says how to call it direct (how letme works).
The same listing from the live API. Graded results come first, then the official MCP registry when no graded-only filter is set.
https://www.anchorterminal.com/api/v1/search
Filters
| Compare | # | Tool | Category | Grade | Score | Agent rating | p95 | Context | Price / x402 | Auth | Where | Details |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 85 | Deepgram Speech-to-Text (Nova-3, Flux)Deepgram · Model API | STT | BB | 76.9 | 3.5 (2) | n/a | n/a | Pay per use | API key | Hosted and local | ||
|
Deepgram's speech-to-text API for recorded audio and live streams, including turn detection for voice agents. Top strength Flux streams with model-level end-of-turn detection, so a voice agent needs no separate VAD Top weakness Training on audio is the default and the opt-out is a per-request flag |
||||||||||||
| 86 | OpenAI Speech to TextOpenAI · Model API | STT | BB | 76.8 | 3.3 (4) | n/a | n/a | Pay per use | API key | Hosted | ||
|
OpenAI's speech-to-text API. It transcribes uploaded audio files through `/v1/audio/transcriptions`, translates recordings into English through `/v1/audio/translations`, and transcribes live audio in Realtime transcription sessions over WebSocket or WebRTC. Top strength
Top weakness
|
||||||||||||
| 96 | Azure AI Speech speech-to-textMicrosoft · Model API | STT | BB | 76.3 | 3.3 (8) | n/a | n/a | Freemium | OAuth or key | Hosted | ||
|
Azure's speech-to-text service for transcribing audio. Top strength Real-time and fast transcription audio isn't stored, and customer audio isn't used for training Top weakness MAI-Transcribe-2 and MAI-Transcribe-2-Streaming are preview with no SLA, and both introductory prices end with 2026 |
||||||||||||
| 110 | Amazon TranscribeAmazon Web Services · Model API | STT | BB | 75.5 | 4.0 (2) | n/a | n/a | Pay per use | API key | Hosted | ||
|
Amazon Transcribe is AWS's speech-to-text API, with batch jobs that read audio from Amazon S3 and real-time streaming over HTTP/2 or WebSocket, signed with IAM credentials. Top strength $0.006 a minute batch and $0.01 streaming in US East, with diarisation, custom vocabulary and language ID included Top weakness AWS may store and use audio to improve the service unless an organisation-wide AI services opt-out policy is set |
||||||||||||
| 111 | ElevenLabs Scribe Speech to Text APIElevenLabs · Model API | STT | BB | 75.5 | 3.5 (2) | n/a | n/a | Freemium | API key | Hosted and local | ||
|
ElevenLabs' speech-to-text service for audio transcription. Top strength $0.22 an hour for batch with 90+ languages and diarisation to 32 speakers Top weakness Audio may be used for training unless the account opts out, and the opt-out isn't retroactive |
||||||||||||
| 209 | Groq Speech-to-TextGroq · Model API | STT | BB | 72.6 | 3.5 (4) | n/a | n/a | Freemium | API key | Hosted | ||
|
Groq's hosted speech-to-text API. It runs OpenAI's Whisper Large v3 and Whisper Large v3 Turbo on OpenAI-compatible transcription and translation endpoints, for uploaded files or audio URLs, with a half-price batch mode. Top strength Published prices of $0.04 an audio hour for Whisper Large v3 Turbo and $0.111 for Whisper Large v3, with 50% off through the Batch API Top weakness No streaming or realtime endpoint and no diarisation in the reviewed documentation |
||||||||||||
| 227 | Gladia Speech-to-Text API + MCPGladia · Model API | STT | BB | 72.3 | 3.0 (2) | n/a | n/a | Freemium | API key | Hosted and local | ||
|
Speech-to-text API for live and recorded audio, with multilingual transcription and code switching. Top strength 100+ languages on Top weakness Self-serve audio is used for model training by default unless the account opts out |
||||||||||||
| 301 | Speechmatics Speech-to-TextSpeechmatics · Model API | STT | BB | 70.3 | 3.5 (2) | n/a | n/a | Pay per use | API key | Hosted | ||
|
Speechmatics' APIs for batch and real-time transcription, including speaker-attributed turns for voice agents. Top strength Training is opt-in only, and realtime audio isn't stored Top weakness Enhanced costs $0.38 to $0.40 an hour, above most rivals |
||||||||||||
| 323 | Soniox Speech-to-TextSoniox · Model API | STT | B | 69.8 | 3.5 (2) | n/a | n/a | Pay per use | API key | Hosted | ||
|
One multilingual model family for 60+ languages, as a real-time WebSocket API (`stt-rt-v6`) and an async file API (`stt-async-v6`). Top strength About $0.10 an hour async and $0.12 real time, with diarisation, language ID and translation included Top weakness No free credits for new accounts since October 2025 |
||||||||||||
| 347 | AssemblyAI Speech-to-Text (Universal)AssemblyAI · Model API | STT | B | 69.3 | 3.5 (2) | n/a | n/a | Pay per use | API key | Hosted | ||
|
Speech-to-text APIs for recorded audio and live streams, with speaker identification, translation and redaction options. Top strength OpenAPI 3.1 file with typed inputs and error responses on every operation Top weakness Fourteen partial incidents from 3 August to 1 November 2026, among them a 75-minute US async outage on 16 September |
||||||||||||
| 351 | Cartesia InkCartesia · Model API | STT | B | 69.2 | 3.3 (4) | n/a | n/a | Freemium | API key | Hosted | ||
|
Cartesia's hosted speech-to-text API. Ink 2 transcribes live audio in five languages over a WebSocket with built-in turn detection, and the older Ink Whisper model transcribes uploaded files in about 100 languages. Top strength
Top weakness The terms let Cartesia train models on inputs and outputs unless otherwise agreed. Opting out is a form in the Playground's data controls |
||||||||||||
| 423 | Google Cloud Speech-to-TextGoogle Cloud · Model API | STT | B | 67.8 | 3.0 (2) | n/a | n/a | Freemium | OAuth | Hosted | ||
|
Google Cloud's transcription API. Top strength Audio isn't stored or used for training unless the project opts in to data logging Top weakness No release note since 2025-11-13 |
||||||||||||
| 442 | Smallest.ai Pulse STTSmallest AI · Model API | STT | B | 67.5 | none | n/a | n/a | Pay per use | API key | Local | ||
|
Smallest AI's Pulse speech-to-text API transcribes pre-recorded files over HTTP and live audio over WebSocket, with three models (Pulse, Pulse 2.0 and Pulse Pro), diarisation, word timestamps, redaction and turn detection. Top strength OpenAPI 3.1 and AsyncAPI 2.6 files cover the batch route and the streaming socket, with llms.txt and Markdown copies of every page Top weakness Health checks on the shared API failed in all three regions on 9, 10, 13, 14 and 15 September 2026, the longest about 64 minutes |
||||||||||||
| open | Parakeet TDT 0.6B v3NVIDIA · Open-weight model | STT | B | 66.9 | 4.0 (4) | n/a | n/a | Free | None | Local | ||
|
Parakeet TDT 0.6B v3 is NVIDIA's open-weight speech recognition model for 25 European languages, a 0.6B-parameter FastConformer encoder with a TDT decoder that detects the language and returns punctuated text with word and segment timestamps, under CC BY 4.0. Top strength Small enough for any machine, 0.71 GB as NVIDIA's Q8_0 GGUF or 2.5 GB in FP32 Top weakness Among the six runtimes only Transformers runs it; vLLM, SGLang, llama.cpp, Ollama and LM Studio don't |
||||||||||||
| 528 | Mistral Voxtral TranscribeMistral AI · Model API | STT | B | 65.6 | 4.0 (4) | n/a | n/a | Freemium | API key | Hosted | ||
|
Mistral AI's speech-to-text API. Voxtral Mini Transcribe 2 transcribes files of up to about three hours with diarisation, word timestamps and context biasing, and Voxtral Realtime transcribes live audio over a WebSocket. Top strength Published prices of $0.003 a minute for Voxtral Mini Transcribe 2 and $0.006 a minute for Voxtral Realtime Top weakness Voxtral Mini Transcribe 2 and Realtime 2602 retire on 30 April 2027, and their Voxtral 3 replacements are Public Preview with silent updates allowed |
||||||||||||
| open | Qwen3-ASR-1.7BAlibaba (Qwen team) · Open-weight model | STT | B | 64.4 | 3.3 (4) | n/a | n/a | Free | None | Local | ||
|
Qwen3-ASR-1.7B is Alibaba's open-weight speech recognition model for 30 languages and 22 Chinese dialects, with language identification, context hints and streaming, published on Hugging Face in BF16 under Apache 2.0. Top strength Apache 2.0 with nothing added, ungated, safetensors only in Qwen's own organisation Top weakness No quantised build from Alibaba, and nothing in Ollama's or LM Studio's library |
||||||||||||
| open | Voxtral Mini 4B Realtime 2602Mistral AI · Open-weight model | STT | B | 62 | 3.5 (4) | n/a | n/a | Free | None | Local | ||
|
Voxtral Mini 4B Realtime 2602 is Mistral AI's open-weight streaming speech recognition model, a causal audio encoder of about 970M parameters on a Ministral 3B decoder, 4.4B in all, transcribing 13 languages live under Apache 2.0. Top strength Natively streaming, with a delay chosen at serving time from 80 ms to 2.4 s Top weakness Plain text out; no timestamps, segments, diarisation or JSON documented for the weights |
||||||||||||
| open | Qwen3-ASR-0.6BAlibaba (Qwen team) · Open-weight model | STT | C | 61.2 | 3.0 (4) | n/a | n/a | Free | None | Local | ||
|
Qwen3-ASR-0.6B is Alibaba's compact open-weight speech recognition model for 30 languages and 22 Chinese dialects, with language identification, context hints and streaming, published on Hugging Face in BF16 under Apache 2.0. Top strength Apache 2.0 with nothing added, ungated, safetensors only in Qwen's own organisation Top weakness No quantised build from Alibaba, and nothing in Ollama's or LM Studio's library |
||||||||||||
| open | Nemotron 3 Nano Omni 30B A3B ReasoningNVIDIA · Open-weight model | Open weights | C | 61.2 | 3.0 (4) | n/a | n/a | Free | None | Local | ||
|
Nemotron 3 Nano Omni is NVIDIA's open-weight multimodal model that takes video, audio, images and text and returns text, a hybrid Mamba-Transformer mixture of experts with about 31B parameters and 3B active, published in BF16, FP8 and NVFP4. Top strength Video, audio, images and text in one model, with word-level timestamps for transcription per the card Top weakness Loads with trust_remote_code in every documented runtime command |
||||||||||||
| open | Granite Speech 5.0 2BIBM · Open-weight model | STT | C | 61 | none | n/a | n/a | Free | None | Local | ||
|
Granite Speech 5.0 2B is IBM's open-weight speech model of 2.3B parameters for transcription in six languages and speech translation to and from English, with keyword lists. Apache 2.0, with IBM's own GGUF builds from about 2 GB. Top strength Apache 2.0, ungated, with a signed repository Top weakness About two minutes of audio per request, so long recordings must be split |
||||||||||||
| open | Nemotron 3.5 Transcribe 30B-A4BNVIDIA · Open-weight model | STT | C | 61 | none | n/a | n/a | Free | None | Local | ||
|
Nemotron 3.5 Transcribe is NVIDIA's open-weight speech model of 30B parameters, about 4.3B active, that transcribes and translates recordings up to an hour in 40 languages, with speaker tags and word timestamps. OpenMDW-1.1. The 4-bit build fits 32 GB. Top strength Transcription, translation, speaker tags and audio questions from one model Top weakness 21.6 GB at 4 bits, so it needs a 32 GB GPU at least |
||||||||||||
| open | Cohere Transcribe 03-2026Cohere · Open-weight model | STT | C | 60.3 | 3.0 (4) | n/a | n/a | Free | None | Local | ||
|
Cohere Transcribe 03-2026 is Cohere's open-weight speech recognition model, a 2B Conformer encoder with a light Transformer decoder trained from scratch, transcribing 14 languages from a pre-specified language, under Apache 2.0, in a gated Hugging Face repository. Top strength Publisher-reported 5.42 average WER on the Open ASR Leaderboard's English sets, with results in the repository to check Top weakness Gated, so an agent needs a Hugging Face account and must accept the conditions before downloading |
||||||||||||
| open | Nemotron 3.5 ASR Streaming 0.6BNVIDIA · Open-weight model | STT | C | 60.3 | 2.8 (4) | n/a | n/a | Free | None | Local | ||
|
Nemotron 3.5 ASR Streaming 0.6B is NVIDIA's open-weight streaming speech recognition model, a 600M-parameter FastConformer with an RNN-T decoder, transcribing 32 language-locales with punctuation and an optional detected-language tag at chunk sizes from 80 ms to 1.12 s, under OpenMDW-1.1. Top strength 0.64B parameters, with a 0.74 GB Q8_0 GGUF from NVIDIA and F32 safetensors of 2.55 GB Top weakness Not in vLLM, SGLang, llama.cpp, Ollama or LM Studio; the GGUF needs NVIDIA's NeMo-Speech.cpp |
||||||||||||
| 898 | Inworld STTInworld AI · Model API | STT | C | 57.2 | none | n/a | n/a | Pay per use | API key | Local | ||
|
Inworld's speech-to-text API runs its `inworld/inworld-stt-1` model in 30 languages over synchronous HTTP, bidirectional WebSocket streaming and asynchronous jobs, returning Voice Profile signals (age, emotion, pitch, accent) and turn detection alongside the transcript. Top strength $0.15 an hour on the free plan and $0.10 on paid plans, billed per second, with 400 free minutes Top weakness No OpenAPI or other machine-readable contract |
||||||||||||
| 907 | IBM Watson Speech to TextIBM · Model API | STT | C | 57 | none | n/a | n/a | Freemium | API key | Local | ||
|
IBM Watson Speech to Text transcribes audio over synchronous and asynchronous HTTP and a WebSocket interface on IBM Cloud, with large speech models, next-generation models, custom language and acoustic models, speaker labels and language identification. Top strength Sync, async with callbacks and WebSocket streaming on one service Top weakness No public OpenAPI file or llms.txt |
||||||||||||
| 917 | Rev AI Speech-to-Text APIRev · Model API | STT | C | 56.7 | 2.5 (2) | n/a | n/a | Freemium | API key | Hosted | ||
|
Rev's API for recorded and streaming speech transcription, with human transcription available through the same job endpoint. Top strength Reverb English at $0.20 an hour, foreign languages at $0.30 Top weakness The streaming WebSocket takes the account's access token as a URL query parameter |
||||||||||||
| open | Ming-3.0-asr-tinyinclusionAI (Ant Group) · Open-weight model | STT | C | 54.8 | 3.0 (4) | n/a | n/a | Free | None | Local | ||
|
Ming-3.0-asr-tiny is inclusionAI's open-weight speech recognition model for Chinese, English and Chinese dialects, an audio encoder on the Ling-tiny-v3 mixture of experts with 8.9B parameters counted from the files, that takes hotwords or dialogue context as hints, under MIT. Top strength MIT with nothing added, ungated, files hashed in inclusionAI's own organisation Top weakness Loading runs the publisher's Python with trust_remote_code, and the reference path needs a CUDA GPU |
||||||||||||
| open | Ming-3.0-audio-tinyinclusionAI (Ant Group) · Open-weight model | STT | D | 52.1 | 2.5 (4) | n/a | n/a | Free | None | Local | ||
|
Ming-3.0-audio-tiny is inclusionAI's open-weight audio-language model for Chinese and English speech recognition, audio understanding and spoken replies, 9.3B parameters with 2.7B active, published on Hugging Face under MIT. Top strength MIT weights in safetensors, downloadable with no account Top weakness No named runtime runs it in a released version, only SGLang-Omni main and inclusionAI's own branches |
||||||||||||
| 1169 | MiniMax Speech-to-TextMiniMax · Model API | STT | D | 49.7 | none | n/a | n/a | Pay per use | API key | Local | ||
|
MiniMax's speech-to-text API transcribes audio files of up to 500 seconds with its `asr-1.0` model in 20 languages, with optional streamed text output, speaker diarisation, segment timestamps and SRT or VTT subtitles. Top strength OpenAPI 3.1 file with eight documented responses, and an error table with HTTP status, type and trigger Top weakness Files only, up to 500 seconds and 50 MB; live audio input isn't supported |
||||||||||||
| open | Parakeet Unified EN 0.6BNVIDIA · Open-weight model | STT | D | 47.5 | none | n/a | n/a | Free | None | Local | ||
|
Parakeet Unified EN 0.6B is NVIDIA's open-weight English speech recognition model, a 600M FastConformer encoder with an RNN-T decoder trained for both offline and streaming use down to 160 ms of latency, with punctuation, published as a NeMo archive. Top strength One model for offline and streaming transcription, down to 160 ms of latency Top weakness The only file is a NeMo archive holding a pickle checkpoint |
||||||||||||
| open | Ming-3.0-audio-flashinclusionAI (Ant Group) · Open-weight model | STT | E | 44.8 | 2.3 (4) | n/a | n/a | Free | None | Local | ||
|
Ming-3.0-audio-flash is inclusionAI's open-weight audio-language model for Chinese and English. It takes audio with text instructions and returns transcripts, answers about the audio, or spoken replies, 126B parameters with 6.5B active, under MIT. Top strength MIT with no use policy, downloadable with no account or gate Top weakness 259.3 GB of BF16 files and no quantised build, so it needs a multi-GPU server and doesn't fit 128 GB |
||||||||||||
| open | Underdog Bark 3.0Conway Research · Open-weight model | STT | E | 44.5 | 2.8 (4) | n/a | n/a | Free | None | Local | ||
|
Underdog Bark 3.0 is Conway Research's open-weight English speech recognition model, a post-trained and compressed Parakeet TDT 0.6B v3 in one 163 MB file, run with Conway's loader on Transformers, MLX or Core ML, under CC BY-NC 4.0. Top strength Small download, 163.4 MB for model.bark and about 0.2 GB for each Core ML package, expanding to about 2.4 GB of FP32 weights Top weakness CC BY-NC 4.0 since 5 November 2026, with no dated note on the card, so commercial use needs a separate licence from Conway |
||||||||||||
| 1359 | Giggy Speech to TextGiggy · Model API | STT | E | 41.4 | none | n/a | n/a | Freemium | API key | Hosted | ||
|
Giggy's hosted speech-to-text API transcribes 16 kHz mono WAV recordings in free Batch, prepaid Fast and Streaming modes, and live audio over a WebSocket, with Python and Node SDKs and a remote MCP server. Top strength Fast is $0.04 and Streaming $0.08 per input audio hour on Pro credits, and Batch is free for recordings of 0.5 seconds to 25 minutes Top weakness Recorded input must be PCM16 mono 16 kHz WAV, so MP3, M4A, stereo or other sample rates need converting before upload |
||||||||||||
Nothing matches these filters. .
p95 latency and context cost come from our probes, which haven't run yet, so those columns start hidden. Grades run from AA to F, and agent-ready means BB or better. Filters, sorting and export run in your browser; the table is complete without JavaScript.
Indexed, not reviewed (19)
Listings sorted into this category from public catalogues (the official MCP registry, APIs.guru, the x402 Bazaar and OpenRouter), with facts and our own checks but no score, grade or rank. How the index works.
| Listing | Kind | What it does | Why it's here |
|---|---|---|---|
| 2kw.ai 2kw.ai | MCP server | EU-hosted AI platform: OpenAI-compatible LLM gateway, document extraction, transcription, agents. | vendor's own, widely used |
| BibiGPT YouTube Video Summarizer bibigpt.co | MCP server | AI video summarizer: summarize and transcribe YouTube, Bilibili, TikTok and podcast videos. | vendor's own |
| clipy.online MCP server clipy.online | MCP server | Read your Clipy screen recordings: search, transcripts, AI summaries, and key moments with frames. | vendor's own |
| DialMCP dialmcp.com | MCP server | Let AI agents place real phone calls from your verified number, with transcripts and recordings. | vendor's own |
| Huddo huddo.ai | MCP server | Group chat for people and AI agents: join a room by invite link, read, reply, whisper, wait. | vendor's own |
| KediMeet kedimeet.app | MCP server | Search and read call transcripts and summaries recorded by the KediMeet Mac app. Read-only. | vendor's own |
| LinkDigest — Xiaohongshu, Douyin, TikTok, YouTube to text linkdigest.dev | MCP server | Douyin, Xiaohongshu, TikTok, YouTube, X links to text: transcript, screen text, viral breakdown | vendor's own |
| Memo AI – meeting assistant memoai.tech | MCP server | Search transcripts and summaries of your meetings, calls, and recordings in Memo AI | vendor's own |
| note1 note1.ai | MCP server | Meeting notes, transcripts, action items with linked tickets, integrations and recordings. | vendor's own |
| SOAPNoteAPI soapnoteapi.com | MCP server | Generate clinical SOAP notes, billing codes, and visit summaries from transcripts or audio. | vendor's own |
| Sosumi sosumi.ai | MCP server | Search and read Apple developer documentation, HIG, and WWDC transcripts as Markdown. | vendor's own, widely used |
| transcriptdock.com MCP server transcriptdock.com | MCP server | YouTube and TikTok transcripts: get_video_transcript in one call, jobs, exports, video discovery. | vendor's own |
| txscribe txscribe.ai | MCP server | Transcribe a link, then read the transcript and summary from your txscribe account. | vendor's own |
| Vexa vexa.ai | MCP server | Meeting bot and transcripts for Google Meet, Teams and Zoom. Live or after, speakers labelled. | vendor's own, widely used |
| Video Extract yanlinglabs.com | MCP server | Download any video from a URL, or get its transcript and key frames. All local, no API keys. | vendor's own |
| Voxplo voxplo.ai | MCP server | Give AI agents a phone: outbound AI calls that return a summary, transcript, and extracted fields. | vendor's own |
| WireCat Zoom wirecat.dev | MCP server | Zoom cloud recordings and transcripts for agents, kept on your computer. Sign in with zm first. | vendor's own, widely used |
| YouTube Transcript, Comments & Video Data API MCP Server youtubeapi.io | MCP server | Real-time YouTube transcripts, comments, channels & video search MCP. Local npx or hosted OAuth. | vendor's own |
| YouTube Transcripts (cloud-safe, via Apify) shephouse-guides.pages.dev | MCP server | YouTube transcripts for agents on cloud servers via an Apify Actor. By ShepHouse (AI-operated). | vendor's own |
How we test this category
The same audio set through every API, with accents, background noise, proper names and overlapping speakers. We measure word error rate on each slice, streaming latency to a final transcript, and cost per audio minute.
How the ranking works
Every tool is scored 0 to 100 across 9 weighted categories, minus deductions for negative events, and given a grade from AA to F. Reliability, latency and context cost are measured by probes and by traffic through letme. Schema quality, security and payments are assessed against a published checklist. Task success comes from a fixed task suite. Payments, including x402, are one category of the nine. Read the methodology.
For companies
Do agents find, use and choose your tools?
An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.